Deep Learning Microsleep Classification

Abstract

Convolutional neural networks (CNNs) are among the deep learning architectures driving the current artificial intelligence revolution and were initially developed for image processing. Image tensors can also be generated from timefrequency analyses such as the continuous wavelet transform (CWT). We investigate whether the combination of CWT and CNN can be successfully applied to the processing of short EEG segments during microsleep (MS) events. Our dataset consists of 24,156 MS and 18,686 counterexamples from five studies conducted in our driving simulator over the past 20 years. When comparing several CNN methods, the ResNet152 network proves to be most successful with accuracies of 96.8 ± 0.9 % after 10-fold cross-validation. This is comparable to the performance of standard machine learning using support vector machines (SVM), which yielded 96.3 ± 0.0 %.